\documentclass[12pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage{graphicx}
\usepackage{setspace}
\usepackage{hyperref}
\usepackage{multirow}
\usepackage{ccaption}
\usepackage{amsmath}
\usepackage{booktabs}
\usepackage{mdwlist}
\usepackage{natbib}
%\bibpunct{(}{)}{;}{a}{}{,} %set in-line reference punctuation
\bibpunct{(}{)}{;}{a}{,}{,} %set in-line reference punctuation
\setlength{\bibsep}{0.1in} %set spacing between references
\setlength{\bibhang}{.1in} %set hanging indent for references
\renewcommand\bibsection{\section*{\refname}} %Changes "Bibliography" to "References"
\expandafter\def\expandafter\quote\expandafter{\quote\singlespacing} %single space quote environment

\captionnamefont{\bfseries}
\captiontitlefont{\bfseries}
\newcommand{\tablenote}[1]{{\singlespacing\footnotesize Note: #1}}

\renewcommand{\footnotesize}{\normalsize}
%\newcommand{\footnoted}[1]{\footnote{\begin{doublespace}\normalsize{#1}\vspace{-26pt}\end{doublespace}}} % 12 pt, double spaced footnotes 
\newcommand{\footnoted}[1]{\footnote{\normalsize{#1}}} % 12 pt, single-spaced footnotes 

\usepackage{tikz}
\usetikzlibrary{shapes,arrows,positioning,decorations.pathreplacing}

\hypersetup{
    unicode=false,          % non-Latin characters in Acrobat's bookmarks
    pdftoolbar=true,        % show Acrobat's toolbar?
    pdfmenubar=true,        % show Acrobat's menu?
    pdffitwindow=false,     % window fit to page when opened
    pdfstartview={FitH},    % fits the width of the page to the window
    pdftitle={Raising the Floor or Closing the Gap? How Media Choice and Media Content Impact Political Knowledge},    % title
    pdfauthor={Thomas J. Leeper},     % author
    pdfsubject={Political Science},   % subject of the document
    pdfkeywords={politics} {public opinion} {political communication} {news attention} {media effects}, % list of keywords
    pdfnewwindow=true,      % links in new window
    pdfborder={0 0 0}
}

\title{Raising the Floor or Closing the Gap? How Media Choice and Media Content Impact Political Knowledge}
\author{}
%\author{Thomas J. Leeper\thanks{Department of Government, London School of Economics and Political Science. The author graciously acknowledges support from NSF Doctoral Dissertation Improvement Grant SES-1160156. This paper also reports survey results found through searches of the iPoll databank and archives of the Roper Center for Public Opinon Research. Data, replication code, and experimental materials will be available on the Harvard Dataverse: \href{https://dataverse.harvard.edu/dataverse/leeper}{https://dataverse.harvard.edu/dataverse/leeper}. Thanks are due to Kevin Arceneaux, Jason Barabas, Matt Baum, Hajo Boomgaarden, Jamie Druckman, Frederik Hjorth, Shanto Iyengar, Jenn Jerit, Matt Levendusky, Paul Marx, Rasmus Tue Pedersen, Markus Prior, and Josh Robison, as well as seminar participants at Aarhus University, the London School of Economics, McGill University, Lousiana State University, Pompeu Fabra University, the European University Institute, and the reviewers and editor at BJPS for helpful feedback on various stages of this project. This paper was previously presented at the 2013 Annual Meeting of the American Political Science Association, Chicago, IL, the 2014 Annual Meeting of the Danish Political Science Association, Vejle, Denmark, and the 2015 Annual Meeting of the Midwest Political Science Association Annual Meeting.}}


\begin{document}

\maketitle

{\abstract\noindent Mass media are frequently cited as having the potential to inform the public, raising knowledge levels and reducing political knowledge gaps between citizens. But media are also seen as a force for segmentation, disengagement, and widening differences between citizens. If media have no effect on political knowledge, gaps between the engaged and disengaged persist regardless of who is exposed to news because no one learns. But gaps can also persist even if everyone learns from the news, particularly if learning effects are heterogeneous across those inclined and disinclined to seek out news and/or across environments that consist of different media alternatives. Yet past research on political communication has not sufficiently linked media choice to debates about possibly heterogeneous effects of media exposure on political knowledge levels. The present study contributes a novel and large-scale choice-based experiment on knowledge of the ongoing crisis in Syria that finds media effects are relatively homogeneous across those with different media preferences and across different media environments. This suggests that under most conditions --- even when everyone learns from the news --- knowledge gaps between the politically engaged and disengaged are widened or at least sustained after incidental exposure to politics. While closing such gaps may be impossible, the results have important implications for understanding how citizens learn about politics and how to study learning from self-selected media experiences.}

\clearpage
\doublespacing

The question of what --- if anything --- citizens know about politics has provoked one of the most central and longest-running debates in the history of political science. That citizens seem to know little creates tensions between researchers' understanding of democratic theory and the empirical realities of contemporary citizenship \citep[see][]{Althaus2006, Berelson1952, DelliCarpiniKeeter1997}. Determining whether citizens know enough about politics, however, necessarily begs the question, inviting circular logics and normative debates about information sufficiency and the definition of competence \citep{Lupia2015}. A more fruitful path tries to answer an empirical question: ``How do citizens become knowledgeable about politics?'' One major factor has been consistently highlighted in the extant literature: exposure to political information via mass media. In particular, policy-specific knowledge of current affairs issues is thought to be highly dependent on media use \citep[see, for a review,][]{Barabasetal2015}.

In these debates, considerable attention is paid to absolute levels of political knowledge; that is, how much citizens know about particular political topics. If knowledge levels are low then understanding how to increase absolute levels of political knowledge is clearly important. A related but distinct question is whether variation in political knowledge levels appears to constitute ``knowledge gaps'' between different groups in society. The politically engaged appear to know much more about politics than those who are less educated or who are politically disengaged. These gaps matter if they influence citizens' opinion formation, vote choices, or other political behaviour \citep[see][]{TichenorDonohueOlien1970}. Knowledge gaps may widen over time and over the course of political debates due to selective (in)attention to political news with those already knowledgeable becoming more so and others learning less. In this view, knowledge gaps are persistently exacerbated by further media exposure. But ``incidental'' or ``byproduct'' media exposure may mean that knowledge gaps are easily lessened or even eliminated by casual exposure to political content in the course of a largely apolitical news diet \citep[see, prominently,][]{Baum2002}. However, this debate is characterized by substantial inferential challenges in disentangling selection biases from the causal influence of media. Given the limitations associated with extant survey and experimental approaches for assessing the influence of media, there is an understandable defeatism epitomized by \citeauthor{Bartels1993}' (1993) famous quip that ``[t]he state of research on media effects is one of the most notable embarrassments of modern social science'' (267). Yet research on knowledge gaps (and media effects more generally) is too important to be abandoned in the face of empirical difficulties. 

This paper suggests that studying media influence on knowledge and other outcomes requires theory and methods that accommodates potentially heterogeneous effects of media across citizens with preferences for different media content. If everyone learns equally from the news, then raising knowledge levels is merely a matter of getting citizens exposed to politics. If, however, learning is heterogeneous and perhaps correlated with preferences over types of media content, then efforts to raise knowledge in absolute terms may come into tension with getting knowledge into the hands of those who need it most. While other sources of heterogeneity are also important, heterogeneity due to differences in media preferences is frequently highlighted as a potential driver of echo chambers and political disengagement (for a review see \citealt{Prior2013}). 

To assess such heterogeneity, the paper presents a choice-focused paradigm for studying mass media influence. Using a ``preference trial'' experiment, which observes \textit{and} manipulates media choices and media content, the study enables a clear assessment of the average effect of media exposure on political knowledge levels as well as potentially heterogeneous effects between those who are inclined and disinclined to seek out political news in different types of media environments. Going beyond previous work, the experimental design positions participants in counterfactual media environments with varying degrees and types of media choices, enabling an assessment of how learning occurs in response to particular content and in the context of particular media choice sets. Specifically, the design uses participants' behavior to reveal their content preferences and, conditional on those preferences and the set of choices available to participants, exposes any heterogeneity in the effects of exposure to news content. Rather than heterogeneous effects, however, I find individuals inclined and disinclined to obtain news learn from it the same amount. Comparisons of various counterfactual conditions further demonstrate that these uniform gains in policy-specific knowledge sustain and even widen gaps between those inclined and disinclined to seek out political content. Indeed, there are no realistic conditions where gaps are overcome by media exposure, raising questions about whether closing such gaps is possible or even a normatively appropriate standard. Ultimately, the findings have important implications for understanding political knowledge, media effects, selective exposure, and the study of media use.


\section*{Media Effects on Political Knowledge}

Political knowledge is widely seen as an important resource for citizens because it shapes opinions, decision-making, and behavior \citep{DelliCarpiniKeeter1997, Barabasetal2015}. It is also fairly uncontroversial to claim that mass media influence citizens' political knowledge (at least knowledge of ``surveillance'' facts; \citealt{Barabasetal2015}). Yet the evidence for the informational value of mass media suffers from theoretical and empirical limitations. Theoretically, this literature tends to attribute variations in knowledge to either (a) attention to different types of media (i.e., ``effects'' of television versus newspapers), or (b) quantitative variations in the amount of participants' exposure to media (i.e., hours of television viewing). While potentially theoretically compelling, both approaches are substantially limited by available data. Self-reported media use is typically the only available operationalization of media exposure and, as such, the key independent variable in analyses of knowledge is subject to various response biases and the substantial measurement error \citep[see, for example,][]{Prior2009a, Prior2009b, Prior2013, DilliplaneGoldmanMutz2012}.

Practical limitations aside, the theoretical focus on media types and quantitative measures of exposure tends to downplay the importance of the actual information the received media convey to their audiences. Given sizeable variations in informational content across media, channels, sources, and articles, theorizing media effects in terms of the quantity of one's exposure --- or exposure to ``types'' of media --- is fundamentally limiting. This focus on quantity of exposure is all the more surprising given that research into the effects of media on other outcomes --- such as opinions, beliefs, issue importance, behaviors, etc. --- has shown that qualitative content variations are incredibly important \citep[see, for example,][]{ChongDruckman2007a, IyengarKinder1987, NelsonClawsonOxley1997, PettyCacioppo1986}. This enormous body of evidence demonstrates unequivocally that content differences matter for a whole host of individual outcomes but knowledge is rarely studied in this way, falling back to correlating self-reported \textit{amounts} of exposure with measures of political knowledge. Research explaining variation in political knowledge can therefore benefit from the theoretical and empirical perspectives offered by the broader media effects literature, where theory focuses on what content is received by different individuals.

That theoretical focus on the effects of content-specific variations similarly invites the use of the same \textit{experimental} rather than \textit{observational} paradigm adopted in the broader media effects literature. A review of studies of political knowledge literature, however, reveals a substantial reliance on cross-sectional observational ``causes of effects'' research designs that involve the regression of a knowledge scale on a set of possible explanatory variables drawn from a single survey \citep[see, for example, ][]{Baum2002, DelliCarpiniKeeter1997, EvelandScheufele2000, JeritBarabasBolsen2006, Prior2007, TichenorDonohueOlien1970}. The limitations of this approach for providing causal inference are now well-known. Selection biases, measurement error in the regressors, and the possibility of ambiguous causal ordering all limit the utility of this approach. As such, we know less about the relationship between media exposure and political knowledge than the volume of studies on the topic might suggest. An apparent cross-sectional correlation between media exposure and knowledge might reflect the causal influence of media on knowledge, the influence of knowledge on media use, both, or even neither.

What is even less understood than the average effect of media on knowledge is whether there is heterogeneity in the effects of media exposure on knowledge. Understanding the size and scope of heterogeneity is critical for evaluating how citizens might learn about politics. For example, differences in knowledge levels might exist simply because those who are political engaged learn more from the same media content (i.e., effect heterogeneity) or because they learn the same amount from that content (i.e., effect homogeneity) but just happen to be exposed more often. The former pattern is one of effect heterogeneity, the latter is one characterized only by self-selection. Amounts of exposure and amounts of learning might both play a role in explaining variation in political knowledge, but extant research has been unable to successfully unpack the difference because observational data never characterize citizens' political knowledge under counterfactual conditions where they receive anything other than their preferred content. This means we cannot understand media effects on political knowledge without carefully attending to both self-selection and potentially heterogeneous effects across those who make distinct media choices within a given media environment.

This argument differs from much existing work. While experimental research documents that citizens are affected by media, such research typically ignores the media selection behavior that is essential for understanding knowledge levels and knowledge gaps \citep[but see][]{DruckmanFeinLeeper2012, Leeper2014, GainesKuklinski2011b, ArceneauxJohnson2012, Stroud2011}. A similar argument has been made about the theory and study of partisan media on voting, issue attitudes, and polarization but with little if any attention to political knowledge \citep[see][]{Levendusky2013, ArceneauxJohnson2012}. All of this work has explored the idea put forth by \citet{Hovland1959} that ``In an experiment the audience on whom the effects are being evaluated is one which is fully exposed to the communication. On the other hand, in naturalistic situations with which surveys are typically concerned, the outstanding phenomenon is the limitation of the audience to those who expose themselves to the communication'' \citep[9]{Hovland1959}. Or, as \citet{BennettIyengar2008} similarly note, ``manipulational control actually weakens the ability to generalize to the real world where exposure to politics is typically voluntary'' (724). While the present research is in the same vein as these choice-focused theories and methods of studying partisan media effects, the experimental tests diverge from this work and the present research focuses on political knowledge rather than attitudinal outcomes. 

I argue that individuals who prefer politics are most likely to engage in media use behaviors that are likely to expose them to political content, while those who prefer other content are likely to engage in media use behaviors that let them avoid political content. The outcomes that result from that exposure therefore depend upon citizens' media choice behavior, the alternatives available to them in the media environment, and the degree of heterogeneity in effects of political news exposure. When individuals can choose what media content they receive, and effects of media are homogeneous, we should expect one of four possible consequences:

\begin{enumerate}
\item \textit{Choices satisfied}: when media choices determine exposure (i.e., those who prefer politics receive it and those who prefer something else do not), knowledge levels increase for some and not for others, raising aggregate knowledge levels but widening knowledge gaps.
\item \textit{All news}: when everyone is exposed to political content, regardless of their media content preferences, aggregate knowledge levels \textit{increase} and any existing knowledge gaps \textit{persist}.
\item \textit{Choices ignored}: when individuals who prefer to avoid politics are unintentionally exposed and when political information is withheld from those who prefer such content, aggregate knowledge levels increase and the knowledge gap \textit{narrows}.
\item \textit{All entertainment}: when no one is exposed to political content, no one learns regardless of their preferences or any existing knowledge.
\end{enumerate}

\noindent If effects are instead heterogeneous, then the story is more complicated. If those inclined to politics learn more than others, then levels increase in (1-3) but gaps between groups widen even in (1-2). Such heterogeneity might emerge because those who self-select into politics have prior knowledge that makes them better able to interpret, integrate, or remember new political information. Conversely, if those inclined to politics instead learn less than others, perhaps because of ceiling dynamics that limit how much more they can learn, then rising knowledge under any circumstance (1-3) mean rising levels overall and narrowing gaps between the groups. Lacking prior evidence it is hard to know which of these preconditions or which of these alternative mechanisms might be at work.

Beyond between-person heterogeneity, there may also be institutional or contextual sources of heterogeneous learning. For example, if media \textit{choices} are conditional --- i.e., depend in the contents of the media choice set --- then these patterns are even more complicated. It may be that learning occurs only when choices are few or when those who would otherwise opt for entertainment find themselves choosing something else because it is unavailable. This dependency of learning not only on citizens' preferences or choices but also on the contents of the media landscape has rarely been studied experimentally but seems critically important. No previous research has disentangled these complex linkages between media self-selection, the contents of the media environment, and the effects of exposure on learning all together. To do so requires a research design that both allows for the expression of ecologically realistic media use behaviors but also manipulates the content that individuals receive in order to causally identify the effects of media exposure conditional on that behavior. 

\section*{A Choice-based Experiment}

If theory dictates that the effects of media exposure on political knowledge are linked to media choices, then empirical research needs to be able to observe selection of content \textit{and} observe effects of exposure to content across those with different choice behaviour, unequivocally distinguishing selection biases from causal effects. While the experimental paradigm of media effects research offers \textit{causal identification}, it says nothing about \textit{selection}. While the observational paradigm of media effects research makes claims about selection, effects, or both, it cannot disentangle them. An increasing number of studies have therefore used choice-based experimental designs to understand the self-selection problem \citep[see][]{GainesKuklinski2011a, GainesKuklinski2011b, Levendusky2012, Levendusky2013, ArceneauxJohnsonMurphy2012, ArceneauxJohnson2012, DruckmanFeinLeeper2012, Leeper2014}. \citet{ArceneauxJohnson2012}, for example, have used a selective exposure paradigm to compare expected aggregate outcomes (e.g., issue attitudes) under counterfactuals of forced exposure and media choice. This design tests for the effect of the \textit{choice set} on the outcome of interest without any need to directly measure what choice participants make among those available. That is, ``choice'' is treated as an experimental condition akin to forced exposure.

A harder problem to solve, however, relates to determining the effect a given media treatment has \textit{conditional on having made a particular choice}. In other words, given an individual expresses a choice behaviorally, what is the effect of the treatment they prefer versus disprefer to receive? This is the problem of choice-related heterogeneity, which extant research largely leaves unresolved. \citet{GainesKuklinski2011b} were the first to apply a choice-based experiment in political science, bringing the method to the attention of scholars of political communication and public opinion. It has since been deployed successfully to study media exposure, for example by \citet{Levendusky2013} who used a design wherein participants were asked for their preference between Fox News, MSNBC, and PBS alternatives and then performs subgroup analysis of a randomized experiment across these self-identified groups of participants. Individuals are therefore differentiated based upon a self-report measure and their preferences were ultimately not respected in the experiment.

The present extends this literature, having been designed in response to earlier literature from outside political science on \textit{patient preference trials} \citep{Sidanietal2009a, Burkeetal2008, Clarketal2008, FloydMoyer2010}. In this design, individuals make choices among alternative treatments from within a finite choice set. Subsequently, individuals are randomly assigned to a particular treatment. For example, a patient chooses between surgical and non-surgical therapies, but is randomly assigned to their actual treatment. Conditional on choice, the randomized treatments expose the necessary counterfactual outcomes for inferring the effects of each treatment among individuals making different choices. The observation of behaviorally expressed choice additionally means that aggregate outcome distributions can also be inferred across different counterfactual realities (e.g., a reality where everyone receives their preferred content versus a reality where everyone regardless of preferences is exposed to political information). The design is the perfect framework for studying media influence on knowledge because it captures and disentangles both theoretically important elements of media effects: self-selection into content (e.g., news channels or stories) and the effects of content on outcomes. The design has spawned an growing methodological literature within political science.\footnote{Whereas previous efforts to understand incidental exposure to political information have used surveys to measure self-reported exposure to the types of programs that might contain incidental political content (e.g., late-night satirical news or soft news programs; \citealt{BrewerCao2006, BaekWojcieszak2009, KimVishak2008, XenosBecker2009}), the patient preference trial clearly disentangles selection from effects. \citet{GainesKuklinski2011a} and \citet{Leeper2017} use this design to assess effects of a short textual vignette on opinions for those inclined and disinclined to select the vignette but (a) focus on opinion outcomes rather than knowledge, and (b) use experimental stimuli that --- like \citet{Levendusky2013} --- that do not respect the participant's content preference. The present design attempts to mitigate the risks of that approach.} Later research (after this study was conducted), has expanded the methodological clarity of the design. For example, \citet{Leeper2017} presents an analysis of how randomized treatment effects average the heterogeneous treatment effects of the typically unobserved groups of treatment choosers and non-choosers. Even more recently, \citet{KnoxYamamotoBaumBerinsky2019} have presented a more detailed methodological treatment of the design, with extensions to multi-value treatments and scenarios where participants inaccurately portray their preferences. Like \citet{Leeper2017}, the focus in the present study is on what \citeauthor{KnoxYamamotoBaumBerinsky2019} et al. refer to as the ``average choice-specific treatment effects'' identified by randomized exposure to treatment within each self-selected arm of the study.


\subsection*{Experimental Design and Procedures}

To test the effects of media choices and media content on political knowledge, I implemented a patient preference trial using a large sample of online participants in the autumn of 2012. The design follows directly from the theorized relationships between media choices, media effects, and knowledge gaps. At the beginning of the experiment, participants were asked some general questions about their media use and then participated in the patient preference trial before finally answering factual questions about an ongoing political issue. Participants were told that the investigators were interested in peoples' reactions to some articles, which the survey would ask about after they finished reading them.

The experiment then involved three manipulations. The first manipulation involves whether individuals were randomly assigned to a set of news stories or whether they had the opportunity to choose what set of articles to read. The second manipulation altered the set of content alternatives available to the participants in the choice-based (preference trial) conditions and the third manipulation altered what issue-relevant content participants received. I discuss each of these manipulations in turn. Figure \ref{fig:conditions} shows the experimental design, conditions and sample sizes.

The first manipulation to either a preference trial or a randomized experiment allows for a comparison of traditional ``captive exposure'' experiments to results of a choice-based experiment, to ensure that the act of expressing a preference over alternative content did not substantially modify subjects' behaviour (an exclusion restriction).\footnote{One concern is that assignment to the preference trial arm of the study had a direct effect on the outcomes. This does not appear to have been the case as the outcome (issue knowledge) did not differ between those receiving the news content in the preference trial ($\bar{x} = 0.49$) compared to those in captive conditions ($\bar{x} = 0.49$, $t=0.22$, $p\le0.83$), nor between those in the preference trial ($\bar{x} = 0.32$) and random assignment ($\bar{x} = 0.33$) conditions who were ultimately assigned to entertainment content ($t=0.29$, $p\le0.78$). Given that the assignment to the preference trial and random assignments arms of the experiment appeared inconsequential, the exclusion restriction does not appear to have been violated.} Those in the randomized conditions were simply instructed they would read some news articles and then were shown either four political news articles or four entertainment articles in sequence. The experience of those in the preference trial is as follows.

\begin{figure}
\caption{Experimental Design, Sequence of Stimuli, and Treatment Group Sizes}\label{fig:conditions}
\includegraphics[width=\textwidth]{Figures/designfigure.pdf}
\tablenote{Participants were randomly assigned to either a preference trial or a randomized experiment. Those in the randomized arm were randomly assigned to either four news articles or four entertainment articles. Those in the preference trial arm were presented with a choice set of either 2 or 3 content types (see main text for description) and asked to choose one of the types to read; they were then randomly assigned to a news treatment article about Syria or an entertainment control article as the third article in a sequence of four article otherwise consistent with their choice. Differences in sample sizes across the arms of the preference trial reflect sample self-selection by participants.}
\end{figure}

The second manipulation involved what set of alternatives each preference trial participant received in their media ``choice set.'' Participants were able to pick what kinds of articles they wanted to read from a randomly displayed set of two or three alternative described as:

\begin{itemize*}
\item ``Political and national government news''
\item ``Human interest stories and lifestyle news''
\item ``Celebrity and entertainment news''
\end{itemize*}

\noindent These alternatives correspond roughly to ``hard,'' ``soft,'' and ``entertainment'' news. A feature not show in Figure \ref{fig:conditions} is that some participants were offered all three alternatives, while the remainder were offered pairs of alternatives (i.e., hard and entertainment; soft and entertainment; hard and soft). The variation in choice sets is a robustness check on the main three-choice condition to ensure that any effects of news content were not an artefact of the choice set.\footnote{Though these choice sets are also stylized, they are meant to represent the ecological conditions that individuals face in a dynamic high--choice landscape and capture institutional variations in the availability of news and entertainment.} This manipulation is doubly useful. First, it allows a test of whether the influence of media depends on the particular choices alternative (to foreshadow: it does not).\footnote{Supplemental Appendix \ref{app:choicesetchecks} further shows that the randomization into different choice sets appeared to have no direct effect on any outcome variable.} Second, it allows for a construct validity test by which it is possible to compare participants' \textit{behaviorally expressed} preferences in the trial (i.e., which of the alternatives they choose to read) to their \textit{stated} interests in different types of news (as expressed on an initial, pretreatment questionnaire). Appendix \ref{app:statedrevealed} shows that there is indeed a strong correspondence between stated preferences and behaviorally revealed opt-in to a particular type of news but this relationship is sensitive to the contents of the choice set. So, while it may be intuitive to think of some individuals as ``news choosers'' and others as ``entertainment choosers,'' that categorization is choice set-dependent. Because this measure of preferences is behaviorally revealed, it is only available for those in the preference trial arm of the study; participants in the randomized arm cannot be separated into these discrete preference groups.

The third and final manipulation randomly assigned participants to read an issue-relevant or a control article. Specifically, participants in every condition were presented with four articles in sequence. Participants in the ``captive'' arm of the study were randomly assigned to one of two streams of content: (1) news, where the third article was issue-relevant, or (2) four entertainment stories. Participants in the captive arm did not behaviorally express a preference over types of content and were simply shown the randomly assigned four articles in sequence.

In the preference trial arm of the study, participants were shown three articles consistent with the choice they indicated and one (the third article in the sequence) that was randomly assigned (again, see Figure \ref{fig:conditions} for a schematic of the sequence). The manipulated story addressed either nonpolitical content (the Academy Awards) or provided participants with information about an important ongoing political issue at the time of the study: conflict in Syria. In every preference trial condition, the three choice-consistent stories were all unrelated to the study's outcome of interest (knowledge of Syria). Two thirds of participants received a randomly assigned article about Syria (e.g., an individual chose ``hard news'' and received four hard news stories, one of which was about Syria) and the remainder were assigned to receive an entertainment article (e.g., an individual chose ``hard news'' and received three hard news stories unrelated to Syria and one article about the Academy Awards). This manipulation provides useful counterfactuals that disentangle selection from effects because each subset of the sample (as defined by behaviourally expressed choices in the preference trial) is observed in alternative states of the world where they do or do not receive information about Syria.

While these three manipulations combine to produce a quite complex experiment, the analysis thereof is actually very simple. We are interested in the effect of exposure to issue-specific news content on political knowledge, relative to a baseline exposure to unrelated entertainment content. In the randomized experiment arm of the study, this means simply comparing the mean levels of knowledge in the ``news'' treatment (Syria) and ``entertainment'' (Academy Awards) conditions. In the preference trial arm of the study, this means comparing mean knowledge levels between individuals incidentally exposed to the news article and those incidentally exposed to the entertainment article, nested by their choice of content and by the choice set to which they were assigned. Thus in each case we are interested in a simple mean-difference in knowledge levels between two groups of participants randomly assigned to receive political news versus entertainment.

The analysis will therefore present an estimate of the effect of issue-specific news exposure for the sample as a whole, based on a comparison of the treatment and control conditions in randomized arm of the trial, and well as nine other estimates of the effect of issue-specific news exposure from the other combinations of choice set and participant news choice in the preference trial arm:

\singlespacing
\begin{itemize}
\item From the choice set of ``hard news'', ``soft news'', and ``entertainment'' content:
\begin{itemize}
\item The subset of those who choose ``hard news''
\item The subset of those who choose ``soft news''
\item The subset of those who choose ``entertainment''
\end{itemize}
\item From the choice set of ``hard news'' and ``soft news'' content:
\begin{itemize}
\item The subset of those who choose ``hard news''
\item The subset of those who choose ``soft news''
\end{itemize}
\item From the choice set of ``hard news'' and ``entertainment'' content:
\begin{itemize}
\item The subset of those who choose ``hard news''
\item The subset of those who choose ``entertainment''
\end{itemize}
\item From the choice set of  ``soft news'' and ``entertainment'' content:
\begin{itemize}
\item The subset of those who choose ``soft news''
\item The subset of those who choose ``entertainment''
\end{itemize}
\end{itemize}

\doublespacing

\noindent Despite the apparent complexities of choices and varying choice sets, it is randomized assignment to the third stimulus article that provides analytic leverage in every case. Because this article is always randomly assigned, the treatment effect can always be estimated as a straightforward comparison of the mean-difference in knowledge between those randomly assigned to the Syria news story and those randomly assigned to the control article about the Academy Awards, conditional on cohice and choice set.\footnote{An alternative analytic approach would be to use estimators defined by \citet{GainesKuklinski2011a}. Results using this approach are included in Supplemental Appendix \ref{app:knowmore} and are substantively and statistically identical to those reported in the body of the paper.} These choice- and choice set-conditional effects reveal any preference-based effect heterogeneity across choice contexts and subgroups of those inclined and disinclined to seek out news media.


\subsection*{Issue Context}

The Syria issue was a reasonable topic at the time of the study and fortuitous given the implications of the continuing Syrian civil war on international and, especially, European politics. From an experimental design perspective, the issue was useful precisely because it was not a topic of major news coverage or public concern in the United States, where the study was conducted. At the time of data collection, violence in Syria had garnered the close attention of less than half of the American public. According to polling by CNN, as of February, 2012, 25\% of the American public indicated believing that the U.S. had a responsibility to intervene in Syria.\footnoted{Answers based upon the question ``Do you think the United States has a responsibility to do something about the fighting in Syria between government forces and anti-government groups, or doesn't the United States have this responsibility?'' from iPoll study USORC.021412B, 2/10--2/13/2012.} By May, that number had increased to 33\%.\footnoted{iPoll study USORC.060612A, 5/29--5/31/2012.} Though no directly comparable survey data is available, by August 29\% were very concerned and 43\% were somewhat concerned about the situation in Syria,\footnoted{Answers based upon the question ``In general, how concerned are you about the situation in Syria--very concerned, somewhat concerned, not very concerned, or not concerned at all?'' from iPoll study USORC.081512, 8/7--8/8/2012.} and 46\% favored U.S. aerial military involvement\footnoted{``Would you favor or oppose the US (United States) and other countries using military airplanes and missiles to try to establish zones inside Syria where the opposition forces would be safe from attacks by the Syrian government?''} but only 32\% favored U.S. use of ground forces.\footnoted{``And would you favor or oppose the US (United States) and other countries using ground troops to try to establish zones inside Syria where the opposition forces would be safe from attacks by the Syrian government?''} Thus, while the American public seemed to be increasingly concerned about the situation in Syria, it had not garnered a large share of the public's attention and less than half of the public was favorable toward U.S. military involvement of any kind. Indeed, as data in Appendix \ref{app:issueattention} show, most of the American public reported (between January and August 2012) not following news about Syria particularly closely and only a tiny fraction of the public indicated that it was the news story they were following most closely. The issue thus serves as an interesting case for studying levels of political knowledge. Given that the Syrian crisis has since come to the forefront of international media attention, the experiment provides an interesting examination of media effects from \textit{before} the issue received major coverage. 


\subsection*{Stimulus Material}

All of the news articles used in the study were modified from recent news coverage and edited to be 800 to 850 words in length. To implement the third manipulation, three treatment articles were created. One that presented the Syria issue by focusing on the politics of rebel groups during the war, one that presented the Syria issue by focusing on the effects of the war on children's schooling, and one that said nothing about Syria (focusing on the schedule for the Academy Awards). The full text of all articles are shown in Supplemental Appendices \ref{app:articles} and \ref{app:articles2}. The two Syria stories articles were modified to mention the same five pieces of information that would be assessed in the outcome knowledge battery. Different versions were created to assess the robustness of the results to variations in the particular treatment articles being used. Because the results for the two articles are largely identical, these conditions are combined (and labelled ``news'') in the analysis. Supplemental Appendix \ref{app:hardsoft} compares the two different conditions, providing full results.

A key feature of the design is that, unlike the intentionally mundane material used in most political communication experiments, all stories including the non-manipulated ``control'' articles were intentionally drawn from contemporary news coverage and edited to catch the attention of participants (being in line with their chosen type of news). One concern with the preference trial design is that participants opting into an entertainment condition might have been ``surprised'' by encountering a political news article and read it either or more less attentively than other participants. Supplemental Appendix \ref{app:articles} contains some validation checks related to reading times for the articles. Conditional on choice, there were no significant differences in reading times between those who received the news treatment article and entertainment control article. 


\subsection*{Outcome Measures and Analysis}

After reading all four articles, participants were asked some general questions about their reactions to the content they read (e.g., ``were the articles too long, too short, or just the right length?''), then answered questions on their knowledge about Syria.\footnoted{The study also measured issue attitudes and attitude certainty. Details on these results are included in Supplemental Appendix \ref{app:attitudes}.} Finally, participants answered some demographic questions and the study concluded.

\begin{table*}
\centering
\caption{Issue Knowledge, by Question}\label{tab:know}
\input{Tables/know.tex}
\end{table*}

Knowledge was measured using five items about Syria and the ongoing conflict, which were scored as correct or incorrect\footnoted{``Don't know'' and blank responses were treated as incorrect. Supplemental Appendix \ref{app:knowmore} reports, among other things, the mean number of ``don't know'' responses to each item. These proportions tended to be high overall, with the mean number of such responses being 1.88.} and then additively scaled and divided by 5 to form a measure of proportion of correct knowledge questions ranging from 0 to 1. Table \ref{tab:know} lists the questions (with correct answers in parentheses) and shows the percentages of all participants correctly answering each of the knowledge questions and the bottom rows reports mean scores and standard deviations on the overall scale. %\footnoted{It is possible that participants simply looked up correct answers on the internet or, conversely, that they would have answered more questions correctly had they been properly incentivized \citep{PriorLupia2008}.} 
The choice of these particular political knowledge items can, of course, be criticized. The goal when selecting items was to obtain a concise battery of surveillance knowledge items of varying ``difficulty.'' Assuaging some concerns, the reported results are also robust to alternative specifications of the knowledge outcome, including item-specific analyses of each knowledge question, analysis of ``don't know'' responses, and an item-response theory (IRT) specification of the analysis (see Supplemental Appendix \ref{app:knowmore}). 

\subsection*{Participants}

Participants in the study were recruited from Amazon Mechanical Turk and paid \$0.75 for their participation.\footnote{This level of pay does not meet current ethical requirements of pay for workers on the Amazon Mechanical Turk platform, who generally see themselves as workers rather than research volunteers, should rightly be paid at least United States federal minimum wage for their time. At the time of the study (2012), Amazon Mechanical Turk was a quite new platform for participant recruitment and this rate of compensation was unfortunately conventional.} A total of 2,221 participants took part in the study, which took about 15 minutes to complete. Implementation took place between October 4 and October 15, 2012.\footnote{Participants were also invited to participate in a follow-up wave of interviewing (time 2) three weeks after completing the first wave with an eye toward understanding the durability of learning over-time.Pparticipants were recontacted using MTurkR \citep{Leeper2013a} via an email with the subject line ``Complete 5-minute follow-up survey for \$.25.'' and body that read as follows: ``Thanks for completing my survey a few weeks ago. Complete a 5-minute follow-up survey (20 questions) to earn a \$.25 bonus. You can complete the survey the link below: \{LinkToStudy\} Thanks so much for your participation! Bonuses will be paid in a few days.'' The recontact rate for the second wave was 62\%, with all responses gathered between October 25 and October 30, 2012. Despite some attrition by time 2, demographics for those participating in both panel waves were similar to the overall sample: 61.7\% were female; 79.5\% were white, 6.0\% were African American, 6.7\% were Asian American, 4.7\% were Hispanic; 53.0\% had college degrees; the median age range was 25-34; 36.5\% were Democrats and 39.8\% were Republicans; 50.6\% identified as liberal; and levels of general political and interest were identical to the those for the whole sample. The key finding for the over-time results is that those who were assigned to entertainment conditions learned about the issue between time 1 and time 2 (or cheated on their answers in the time 2 responses), while those assigned to news conditions retained their knowledge over time. Full results for time 2 are reported in the Supplemental Appendix.}
While some have raised concerns about the validity of MTurk, an increasingly large body of evidence suggests the relatively diverse set of participants recruited from the platform \citep{BerinskyHuberLenz2012, HuffTingley2015} experience causal dynamics comparable to other convenience and representative samples \citep{Mullinixetal2015, KrupnikovLevine2014}. Though in no way representative of the U.S. population, participants constituted a diverse sample: 60.4\% were female; 78.5\% were white, 7.1\% were African American, 6.2\% were Asian American, 5.0\% were Hispanic; 48\% had college degrees; median age range was 25-34; 35.6\% were Democrats and 38.0\% were Republicans; 49.6\% identified as liberal. Participants expressed moderate levels of interest in politics (mean=0.6, sd=0.3, on a 0-1 scale). As in any research, the use of a convenience sample --- like the choice of a focal issue and research setting --- may of course limit the generalizability of the results across settings, persons, issues, or outcomes.



\section*{Results}

Receiving mediated information about Syria increased knowledge about the issue. Scores on the knowledge scale were compared across participant groups that received the news and entertainment treatments. In the randomized arm of the experiment, treatment group means were 0.49 (SE=0.01) for the news condition and 0.32 (0.03) for the entertainment condition, a difference that indicates a sizeable and statistically significant ($p=0.00$) gain in issue-relevant political knowledge among those exposed to political news. Receiving information about Syria clearly increases knowledge.

But how are knowledge levels (and the gaps between them) affected by the interaction between media preferences, the media choice landscape, and exposure? Table \ref{tab:knowchoice2} shows mean levels of issue knowledge among individuals assigned to the treatment news article or the control entertainment article, separately by the choice set and, within each choice set, those choosing entertainment, soft news, or hard news.\footnote{Note: those in captive conditions were not given an option to behaviorally express a preference for a type of content, therefore it is not possible to divide them based upon such a preference.} For example, the first three rows represent participants in the preference trial arm of the study presented with three alternatives in the choice set. The first row represents the subset of these participants that expressed a preference for ``hard news'' --- the penultimate column is the mean level of knowledge among these participants when randomly assigned to a news article about Syria and the final column is the mean level of knowledge among these participants when randomly assigned to the control article about the Academy Awards. The next two rows are for the subsets of participants in this same choice set that behaviorally preferred ``soft news'' and ``entertainment,'' respectively. The subsequent rows of the tables represent the same values for each of the two-alternative choice sets.

The highest levels of knowledge were displayed by participants who chose ``hard news'' when it was available and subsequently received issue-specific information (penultimate column of rows 1, 6, and 8). Yet these participants also had high levels of issue-specific knowledge even when they received no issue-relevant content in the experiment (see last column). By contrast, participants who chose entertainment when it was available had the lowest levels of knowledge (last column of rows 3, 5, and 7) --- a clear knowledge gap --- but displayed significantly higher knowledge when they were assigned the treatment article in the experiment (see penultimate column). Recall that individuals in the preference trial were given a choice between \textit{general} categories of news content to read but were assessed on \textit{issue-specific} knowledge, so these differences should be understood as a preference-related knowledge gap rather than a self-selection bias wherein those more knowledgeable about Syria specifically opted to read news about Syria. Even after receiving the Syria story, however, those choosing entertainment know the least, those choosing soft news know somewhat more, and those choosing hard news know the most.

%0.17 ($t$=5.75, $p$<0.00) % chose hard
%0.19 ($t$=8.01, $p$<0.00) % chose soft
%0.15 ($t$=5.84, $p$<0.00) % choice ent

\begin{table*}
	\caption{Mean Issue Knowledge Level by Choice Set, Choice, and Randomized Treatment}\label{tab:knowchoice2}
	\begin{center}
		\input{Tables/knowmeans.tex}
	\end{center}
	{\tablenote Cell entries are treatment group means levels of knowledge for participants in the issue-specific news treatment condition versus the control condition, within each choice set and revealed preference group.}
\end{table*}

Despite baseline differences in knowledge across choice groups, it appears that everyone learned from news exposure. These treatment effects are shown in Figure \ref{fig:effects}, separately for each choice and choice set group. Rather than the opportunity for media choice segmenting those who are engaged and capable of learning from those who are disengaged and learn less, it appears that anyone can learn (and learn roughly the same amount) if given at least an incidental exposure to news.\footnoted{Noteworthy, however, is that by two to three weeks after exposure, the effects of news also dissipates uniformly across all types of individuals in all conditions because those assigned to the entertainment story showed a slight increase in knowledge over the post-treatment period. See Supplemental Appendix \ref{app:knowmore} for full results. This may reflect participants search for information outside of the experiment \citep[see][]{DruckmanFeinLeeper2012}, or possibly some cheating behavior on the part of participants. Unfortunately, it is not possible to test either of these possibilities.} Knowledge \textit{levels} increased for everyone, but these increases mean that knowledge \textit{gaps} persisted.

\begin{figure*}
	\caption{Treatment Effects on Knowledge Levels, by Choice Set and Choice}\label{fig:effects}
	\includegraphics[width=\textwidth]{Figures/effects-by-choiceset1a}
	{\tablenote Figure displays average treatment effects, with bars for 67\% and 95\% confidence intervals, of news exposure (relative to entertainment exposure) in each self-selected choice group separately for each choice set. The vertical brackets represent choice sets, with groups within each bracket representing subsets of participants making different media choices.}
\end{figure*}


How should we normatively interpret these results? One way is to consider a trade-off between satisfying citizens' media preferences (by providing them with their preferred content) and the democratic goal of improving total political knowledge. To do so, we can examine knowledge levels and the knowledge gap between the politically engaged and disengaged across various counterfactual conditions revealed through the experiment's interaction of choices and content provision. For example, what would aggregate knowledge be if those who preferred news received news while those who prefer entertainment received entertainment (``choices satisfied'') versus a scenario where all were forced to read the news? What would knowledge be if all were instead forced to read entertainment? Or, more controversially, if preferences were ignored such that those who prefer news received entertainment and those who prefer entertainment were given news? Comparing knowledge \textit{levels} and the knowledge \textit{gap} (between news choosers and entertainment choosers) in each of these counterfactuals tells us about what the impact of media are in worlds where media preferences are satisfied by the media landscape to varying degrees.

Figure \ref{fig:counterfactuals} shows these four counterfactuals by taking the mean knowledge of those who prefer ``hard news'' and ``entertainment'' when their choices are either satisfied (news choosers get news and entertainment choosers get entertainment) or ignored (everyone receiving the opposite of their preferred content), or when both groups receive news, or when both groups receive entertainment. The horizontal axis shows the aggregate mean level of knowledge for all participants (how knowledgeable would this society be?) and the vertical axis shows the size of the gap in knowledge between those who prefer news and those who prefer entertainment (how unequal would knowledge be in this society?).

\begin{figure*}
	\caption{Knowledge Levels and Knowledge Gaps in Several Counterfactuals}\label{fig:counterfactuals}
	\begin{center}
	\includegraphics[width=\textwidth]{Figures/knowledge_tradeoff1.pdf}
	\end{center}
	{\tablenote Figure summarizes counterfactual conditions extracted from the experimental data wherein individuals do or do not receive their preferred content. The x-axis shows the average knowledge levels in each counterfactual reality and the y-axis shows the mean-difference in knowledge between those inclined to choose political news versus entertainment. Individuals choosing soft news are ignored in these calculations.}
\end{figure*}

When preferences are satisfied, a large gap in knowledge exists because those who choose news and those who prefer entertainment --- levels rise for the former and persist for the latter. Yet if citizens receive their preferred content, knowledge gaps (at least immediately) are likely to widen even if some citizens learn.

The citizenry in a world where everyone is exposed to news is, unsurprisingly, more knowledgeable on average. Yet the public is on average only slightly more knowledgeable in this condition than when only news choosers receive news. By contrast, when everyone receives entertainment instead of news (regardless of their preferences), a knowledge gap persists and overall knowledge is very low.

The final counterfactual is especially interesting: when preference are ignored --- news choosers get entertainment and entertainment choosers get news --- the knowledge gap is almost entirely closed. But this narrowed knowledge gap comes at the expense of the average level of knowledge in the group as a whole; the public knows only slightly more on average than if their media preferences were satisfied. This is the only condition in which political knowledge gaps are substantially narrowed. Yet such a world is hardly democratically palatable. What is better: a world where political knowledge is equitable and low or a world where it is inequitable and high?

If knowledge is essential for civic participation \citep[e.g., as argued by ][]{DelliCarpiniKeeter1997}, then unequal knowledge means unequal political power, something that is hardly desirable. Yet if low-knowledge citizens are able to obtain accurate cues from better-informed peers \citep[see, for example,][]{Lupia1994} and the distributions of political interests (or values) and knowledge in society are independent, then the informationally unequal society may be normatively acceptable. Rather than resolve this debate, this present research even further complicates it by highlighting the challenging connections between the provision of information by a media landscape and the internalization of that information as knowledge by a diverse citizenry.


\section*{Discussion}

The present study shows that the effects of exposure to political news appear to be homogeneous across individuals with different media preferences and across different media choice environments. Media exposure can increase knowledge level in the aggregate as well as increase issue-specific knowledge among those inclined and disinclined to receive that content, at least on the issue of conflict in Syria. What then explains variations in political knowledge levels? It appears that self-selection rather than differential effects is the answer. At the same time, the learning that does occur does not appear to depend on the particular set of alternatives available in the choice set or preferences expressed --- if individuals are exposed to politics, regardless of how they come upon it, they will learn. While the results are derived from a particular sample, at a particular point in time, on a particular issue, they have important implications for our understanding of political knowledge, political communication effects, and the design of experiments.

The study shows that increasing political knowledge through media exposure requires either (a) forced exposure to political news, (b) an inclination on the part of citizens to view political content, or (c) incidental exposure to political news for those who would prefer to see something else also appears to be effective. All are paths to increasing knowledge levels. These processes, however, do not produce equivalent impacts on aggregate knowledge or knowledge gaps. The most realistic of these conditions --- where individuals receive their preferred content --- means modestly high overall knowledge but a widening knowledge gap between those with preferences for news content and those for preferences for other content. The conditions where politics avoiders are incidentally exposed to politics also mean rising knowledge but a sustained (not narrowed) knowledge gap. It may be the case that other political issues would display different patterns of self-selection and learning, but the Syrian case is typical in the sense that it was relatively unfamiliar, relatively new at the time of the study, involved issues relatively distant to the daily lived experience of most Americans, and is similar to that used in previous research on issue-specific knowledge \citep{Baum2002}. Future replications of this study might explore the contextuality of learning across issues and across different media choice sets.

A key result of the study is that for this kind of issue, it is only in the unlikeliest of circumstances --- where individuals only receive their \textit{dispreferred} content --- that political knowledge gaps are narrowed. That equalizing knowledge comes at the expense of learning for those already politically inclined raises significant normative questions about academic debate surrounding political knowledge. While equalizing knowledge might appear desirable, it is not clear that such a goal is more important than simply increasing aggregate knowledge. Similarly, while forcing everyone to view news would be an effective strategy for raising the floor of low knowledge levels, it would not necessarily close gaps given pre-existing differences in knowledge. As Figure \ref{fig:counterfactuals} demonstrated, the normative tension between these two seemingly related goals merits further consideration. 

Speaking to media effect literature more broadly, the present research contributes further evidence that media effects occur as an interaction between individual choices and content provision \citep{Prior2007, Stroud2011}. While much past research has been attentive to either the factors that explain different practices of media use \textit{or} to differences in the informational content provided by different outlets, media, and environments, greater attention should be paid to how these factors interact to impact politically relevant outcomes. Consequently, scholars and commentators concerned with media effects should focus much more attention on the information provided by different media systems \citep[see, for example,][]{AalbergvanAelstCurran2010}. Individual citizens cannot be blamed for following their preferences toward particular content, but the media system can be structured in such a way as to provide useful political information even to those avoiding political news. Similarly, given this interplay between choices and institutions, more attention should be paid to the drivers of selective exposure aside from partisanship, ideology \citep{Stroud2011, ArceneauxJohnsonMurphy2012, ArceneauxJohnson2012, Levendusky2012, Levendusky2013}, or entertainment tastes \citep{Prior2007}.

In terms of experimental design, this study reinforces the value of studying selection into media and the effects of media together. Captive exposure experiments have provided credible but perhaps unrealistic estimates of media effects without accounting for media selection \citep[see, for reviews,][]{NelsonBrynerCarnahan2011, Kinder1998, Kinder2003a}, survey-based analyses show either selection or effects but cannot disentangle influence from selection bias. Empirical research on media effects must therefore move toward directly observing (and directly manipulating) these media preferences and behaviors in order to avoid the causal ambiguity that plagues the literature, following the lead of a growing body of evidence \citep{GainesKuklinski2011b, GainesKuklinski2011a, DruckmanFeinLeeper2012, Feldmanetal2013, Leeper2014, Leeper2017, KnoxYamamotoBaumBerinsky2019}. Future work might apply these methods in other contexts and begin to incorporate other complexities of the current media landscape, like social recommendation \citep{MessingWestwood2014} and content provision through choice-tailored algorithms \citep{BoczkowskiMitchelstein2013, KnoblochWesterwicketal2005}.

More than sixty years ago, \citet{Downs1957} pointed out that ``Society's free information stream systematically provides some citizens with more politically useful information than it provides others'' (221). Some are exposed to politics because they choose it, others because of who they know and where they work encounter politics indirectly, and others live in situations that expose them to politics only infrequently. \citep[222--23]{Downs1957}. Americans at the time of this study knew little about the conflict in Syria, but regardless of their taste for political or entertainment news, it was possible for them to learn a substantial amount about the issue through a relatively brief and perhaps incidental exposure. Yet we still do not know enough about how preferences and choices interact with the media environment and its information provision to produce media effects. Choice-based experiments offer a clear way forward.


\clearpage
\bibliographystyle{apsa-leeper}
\bibliography{references}

\clearpage
\appendix
\singlespacing

\input{SupplementaryMaterials}
\end{document}
